This commit is contained in:
hailin 2025-07-27 18:34:25 +08:00
parent 900be3e02d
commit 79abd2bbdd
1 changed files with 23 additions and 23 deletions

View File

@ -3,11 +3,11 @@ from pathlib import Path
from collections import deque from collections import deque
import queue, threading, time import queue, threading, time
# ───────────────────── 基础配置 ───────────────────── # ────────────────── 基础配置 ──────────────────
API_URL = "http://localhost:30000/generate"
API_KEY = "token-abc123" API_KEY = "token-abc123"
MODEL_PATH = Path("/root/.cradle/Alibaba/Qwen3-30B-A3B-Base") MODEL_PATH = Path("/root/.cradle/Alibaba/Qwen3-30B-A3B-Base")
def model_name(path: Path): def model_name(path: Path):
cfg = path / "config.json" cfg = path / "config.json"
if cfg.exists(): if cfg.exists():
@ -18,18 +18,19 @@ def model_name(path: Path):
MODEL_NAME = model_name(MODEL_PATH) MODEL_NAME = model_name(MODEL_PATH)
now = lambda: datetime.datetime.now().strftime("%H:%M:%S") now = lambda: datetime.datetime.now().strftime("%H:%M:%S")
# ───────────────────── 日志队列 ───────────────────── # ────────────────── 日志队列 ──────────────────
LOG_Q: "queue.Queue[str]" = queue.Queue() LOG_Q: "queue.Queue[str]" = queue.Queue()
LOG_TXT = "" # ✅ 全局日志缓存,避免 chat 焦点阻断 log_box 更新 LOG_TXT = ""
def log(msg): # 写终端 + 推队列
def log(msg):
print(msg, flush=True) print(msg, flush=True)
LOG_Q.put(msg) LOG_Q.put(msg)
prev_log_value = "" # 上一帧的日志内容
prev_log_value = ""
def consume_logs(dummy=None): def consume_logs(dummy=None):
"""每秒更新 log_box 内容,避免 chat 阻塞 UI 刷新"""
global LOG_TXT, prev_log_value global LOG_TXT, prev_log_value
buf = deque(LOG_TXT.splitlines(), maxlen=400) buf = deque(LOG_TXT.splitlines(), maxlen=400)
while not LOG_Q.empty(): while not LOG_Q.empty():
@ -38,10 +39,10 @@ def consume_logs(dummy=None):
if LOG_TXT != prev_log_value: if LOG_TXT != prev_log_value:
prev_log_value = LOG_TXT prev_log_value = LOG_TXT
return gr.update(value=LOG_TXT) return gr.update(value=LOG_TXT)
return gr.update() # 无更新则不触发前端刷新 return gr.update()
# ───────────────────── 后端调用 ───────────────────── # ────────────────── 后端调用 ──────────────────
def backend(text, sampling, api_suffix): def backend(text, sampling, api_suffix):
url = f"http://localhost:30000{api_suffix}" url = f"http://localhost:30000{api_suffix}"
if api_suffix == "/generate": if api_suffix == "/generate":
@ -69,7 +70,7 @@ def backend(text, sampling, api_suffix):
meta = data.get("meta_info", {}) meta = data.get("meta_info", {})
fr = meta.get("finish_reason") fr = meta.get("finish_reason")
ctok = meta.get("completion_tokens") ctok = meta.get("completion_tokens")
else: # "/v1/chat/completions" else:
choice = data.get("choices", [{}])[0] choice = data.get("choices", [{}])[0]
txt = choice.get("message", {}).get("content", "").strip() txt = choice.get("message", {}).get("content", "").strip()
fr = choice.get("finish_reason") fr = choice.get("finish_reason")
@ -85,18 +86,21 @@ def backend(text, sampling, api_suffix):
return f"[❌ 请求异常] {e}" return f"[❌ 请求异常] {e}"
# ───────────────────── Chat 回调 ───────────────────── # ────────────────── Chat 回调 ──────────────────
def chat( def chat(
user, history, user_msg, history,
max_new, temp, top_p, top_k, max_new, temp, top_p, top_k,
rep_pen, pres_pen, stop_raw, rep_pen, pres_pen, stop_raw,
api_suffix, log_state api_suffix, log_state
): ):
from queue import Queue, Empty from queue import Queue, Empty
# 解析传入的 ChatInput 格式
user = user_msg["text"] if isinstance(user_msg, dict) and "text" in user_msg else user_msg
stop = [s.strip() for s in stop_raw.split(",") if s.strip()] or None stop = [s.strip() for s in stop_raw.split(",") if s.strip()] or None
samp = { samp = {
("max_tokens" if api_suffix == "/v1/completion" else "max_new_tokens"): int(max_new), ("max_tokens" if api_suffix == "/v1/chat/completions" else "max_new_tokens"): int(max_new),
"temperature": temp, "temperature": temp,
"top_p": top_p, "top_p": top_p,
"top_k": int(top_k), "top_k": int(top_k),
@ -105,7 +109,6 @@ def chat(
**({"stop": stop} if stop else {}) **({"stop": stop} if stop else {})
} }
result_q = Queue() result_q = Queue()
def worker(): def worker():
@ -123,14 +126,13 @@ def chat(
continue continue
# ───────────────────── Gradio UI ───────────────────── # ────────────────── Gradio UI ──────────────────
with gr.Blocks(title="调试界面") as demo: with gr.Blocks(title="调试界面") as demo:
gr.Markdown(f"## 💬 调试界面 \n权重 **{MODEL_PATH.name}**") gr.Markdown(f"## 💬 调试界面 \n权重 **{MODEL_PATH.name}**")
with gr.Row(): with gr.Row():
api_choice = gr.Dropdown(choices=["/generate", "/v1/completions"], api_choice = gr.Dropdown(choices=["/generate", "/v1/chat/completions"],
value="/generate", label="选择推理接口") value="/generate", label="选择推理接口")
# 采样参数控件
with gr.Row(): with gr.Row():
max_new = gr.Slider(32, 32768, 128, label="max_new_tokens") max_new = gr.Slider(32, 32768, 128, label="max_new_tokens")
temp = gr.Slider(0, 1.5, 0.8, step=0.05, label="temperature") temp = gr.Slider(0, 1.5, 0.8, step=0.05, label="temperature")
@ -142,21 +144,19 @@ with gr.Blocks(title="调试界面") as demo:
pres_pen= gr.Slider(0, 2, 0.0, step=0.05, label="presence_penalty") pres_pen= gr.Slider(0, 2, 0.0, step=0.05, label="presence_penalty")
stop_txt = gr.Textbox("", label="stop 序列(逗号分隔)") stop_txt = gr.Textbox("", label="stop 序列(逗号分隔)")
log_state = gr.State("") # 状态透传 log_state = gr.State("")
dbg_chk = gr.Checkbox(label="📜 显示 Debug 面板", value=False) # ✅ 默认关闭 dbg_chk = gr.Checkbox(label="📜 显示 Debug 面板", value=False)
log_box = gr.Textbox(label="实时日志", lines=20, interactive=False, visible=False) # ✅ 默认隐藏 log_box = gr.Textbox(label="实时日志", lines=20, interactive=False, visible=False)
# Chat 界面(移到日志之前)
chatbot = gr.ChatInterface( chatbot = gr.ChatInterface(
fn=chat, fn=chat,
additional_inputs=[max_new, temp, top_p, top_k, additional_inputs=[max_new, temp, top_p, top_k,
rep_pen, pres_pen, stop_txt, rep_pen, pres_pen, stop_txt,
api_choice, log_state], api_choice, log_state],
additional_outputs=[], # ✅ 移除 log_state 输出 additional_outputs=[],
type="messages" type="messages"
) )
# 日志刷新定时器
timer = gr.Timer(1.0, render=True) timer = gr.Timer(1.0, render=True)
timer.tick( timer.tick(
fn=consume_logs, fn=consume_logs,